Image generators come in two broad families. Diffusion models refine a whole picture in parallel over a number of steps; autoregressive models write it out token by token, the way a language model writes text. The second approach is attractive because it can share one architecture with chat models, but it has been slow, since every extra token is another sequential step. G-STAR, a preprint from Peking University, Meituan's Longcat team and several other groups, attacks that speed problem directly.
The obvious fix is to compress the image harder so there are fewer tokens to generate. The catch is that each token then has to carry more information, and discrete codes lose detail. G-STAR represents each spatial token with a group of codes rather than one, and the authors show that to keep the same information budget, the group size should grow with the square of the compression factor along each side. That lets them push spatial compression to 32 times, and effectively 64 times, well past the common 16. A lightweight decoder predicts most of each group in parallel and keeps only a final causal step, so bigger groups add little delay.
The team trained two models, GSTAR-Pro-8B at 32 times compression and GSTAR-Flash-1.7B at 64 times. At 1024 by 1024 pixels they report a 20-fold speed-up over earlier discrete autoregressive models, about 2.5 seconds per image on a single Nvidia H200, with scores competitive with leading systems on five text-to-image benchmarks: GenEval, DPG-Bench, PRISM, LongText-Bench and Qwen-Image-Bench. In a controlled comparison, moving from 32 to 64 times compression cost only a little quality, with GenEval slipping from 0.8060 to 0.7885. Code and weights are on GitHub and Hugging Face.
This matters for unified models that both understand and generate images, a direction several labs are pursuing because one network can then read a picture, reason about it and edit it. Speed has been the practical barrier to shipping such systems to users who expect an image in seconds.
Read the claims carefully. Competitive is not the same as leading, and the 20-fold gain is measured against older autoregressive models, not the fastest diffusion systems people use today. Benchmark scores for image quality only partly match what users see, and the paper's comparisons were run by the authors. It is a preprint, not peer reviewed, and real-world latency will depend on hardware far less generous than an H200.